In Humphreys Lie Algebra text, before performing root space decomposition it is required to pick a maximal toral subalgebra of the semisimple lie algebra in question. By maximal, he means not properly contained in any other toral subalgebra.
Why must the toral subalgebra $\mathfrak{h}$ be maximal? What fails if we just find an arbitrary semisimple element $x$ of the lie algebra and set $\mathfrak{h}={\rm span}\{x\}$?
What we need often is that the maximal total subalgebra is self-normalising, i.e., that it agrees with the Cartan subalgebra in the semisimple case. If we do not have this, then we do not have a Cartan subalgebra. A trivial example would be to take $\mathfrak{h}=0$, as $0$ is a toral subalgebra (it contains no nonzero nilpotent elements).